Expand calibration samples and extend training retention

artifacts/probability_calibration/default.json

@@ -1,20 +1,20 @@
 {
-  "version": "emos-20260415164136",
+  "version": "emos-20260415170349",
-  "trained_at": "2026-04-15T16:41:36.704829+00:00",
+  "trained_at": "2026-04-15T17:03:49.122865+00:00",
   "global": {
     "mu": {
-      "intercept": -0.39094186,
+      "intercept": 2.45853035,
-      "raw_mu_coef": 0.27476248,
+      "raw_mu_coef": 2.03689031,
-      "deb_coef": 0.91159557,
+      "deb_coef": -1.41207502,
-      "ens_median_coef": -0.16959018,
+      "ens_median_coef": 0.08616811,
-      "max_so_far_gap_coef": -0.94232523
+      "max_so_far_gap_coef": -0.5886898
     },
     "sigma": {
-      "intercept": -1.56485068,
+      "intercept": 0.61260118,
-      "raw_sigma_coef": 0.12930442,
+      "raw_sigma_coef": -1.48747774,
-      "spread_coef": 0.15947819,
+      "spread_coef": 2.24562634,
-      "peak_flag_coef": 0.19328621,
+      "peak_flag_coef": -0.15906747,
-      "max_so_far_gap_coef": -0.08469752
+      "max_so_far_gap_coef": -0.01835251
     }
   },
   "sigma_constraints": {
@@ -29,56 +29,81 @@
     "max_bucket_brier_increase": 0.05
   },
   "blending": {
-    "alpha_mu": 1.0,
+    "alpha_mu": 0.05,
-    "alpha_sigma": 0.5
+    "alpha_sigma": 1.0
   },
   "cities": {
     "tokyo": {
-      "samples": 4,
+      "samples": 5,
-      "mu_bias": 0.211244,
+      "mu_bias": 2.223946,
-      "sigma_scale": 0.5,
+      "sigma_scale": 2.0,
-      "confidence": 0.5
+      "confidence": 0.625
     },
     "ankara": {
-      "samples": 11,
+      "samples": 6,
-      "mu_bias": 0.027471,
+      "mu_bias": -0.396953,
-      "sigma_scale": 0.5,
+      "sigma_scale": 1.094474,
-      "confidence": 1.0
+      "confidence": 0.75
     },
     "istanbul": {
-      "samples": 12,
+      "samples": 3,
-      "mu_bias": 0.048913,
+      "mu_bias": -0.163153,
       "sigma_scale": 0.5,
-      "confidence": 1.0
+      "confidence": 0.375
     },
     "hong kong": {
-      "samples": 12,
+      "samples": 7,
-      "mu_bias": -0.026823,
+      "mu_bias": 3.267553,
-      "sigma_scale": 1.020494,
+      "sigma_scale": 2.0,
-      "confidence": 1.0
+      "confidence": 0.875
     },
     "taipei": {
       "samples": 5,
-      "mu_bias": 0.192719,
+      "mu_bias": 4.80044,
-      "sigma_scale": 0.5,
+      "sigma_scale": 2.0,
       "confidence": 0.625
+    },
+    "seoul": {
+      "samples": 3,
+      "mu_bias": -3.080962,
+      "sigma_scale": 1.437353,
+      "confidence": 0.375
+    },
+    "milan": {
+      "samples": 3,
+      "mu_bias": -4.20255,
+      "sigma_scale": 2.0,
+      "confidence": 0.375
+    },
+    "shanghai": {
+      "samples": 3,
+      "mu_bias": 1.185335,
+      "sigma_scale": 1.68386,
+      "confidence": 0.375
+    },
+    "warsaw": {
+      "samples": 3,
+      "mu_bias": -1.574246,
+      "sigma_scale": 1.343402,
+      "confidence": 0.375
     }
   },
   "metrics": {
-    "sample_count": 50,
+    "sample_count": 71,
-    "mean_crps": 0.278798,
+    "mean_crps": 3.226613,
-    "legacy_mean_crps": 1.26525,
+    "legacy_mean_crps": 3.608885,
-    "legacy_mean_mae": 1.4674,
+    "legacy_mean_mae": 3.81507,
-    "legacy_bucket_hit_rate": 0.54,
+    "legacy_bucket_hit_rate": 0.492958,
-    "legacy_bucket_brier": 0.744518,
+    "legacy_bucket_brier": 0.873776,
-    "selected_mean_crps": 0.278798,
+    "selected_mean_crps": 3.44742,
-    "selected_mean_mae": 0.248457,
+    "selected_mean_mae": 3.749078,
-    "selected_bucket_hit_rate": 0.82,
+    "selected_bucket_hit_rate": 0.492958,
-    "selected_bucket_brier": 0.369542,
+    "selected_bucket_brier": 0.878621,
-    "selected_score": 0.8508,
+    "selected_score": 5.241857,
-    "legacy_score": 2.660378,
+    "legacy_score": 5.406287,
     "filled_actual_from_history": 0,
-    "settlement_history_city_count": 30
+    "settlement_history_city_count": 30,
+    "legacy_archive_samples": 54
   },
   "source": "artifacts\\probability_calibration\\default.json"
 }

artifacts/probability_calibration/evaluation_report.json

@@ -1,95 +1,284 @@
 {
   "summary": {
-    "sample_count": 17,
+    "sample_count": 71,
     "filled_actual_from_history": 0,
     "legacy": {
-      "mean_crps": 1.095786,
+      "mean_crps": 3.608885,
-      "mean_mae": 1.243529,
+      "mean_mae": 3.81507,
-      "bucket_hit_rate": 0.588235
+      "bucket_hit_rate": 0.492958
     },
     "emos": {
-      "mean_crps": 1.257578,
+      "mean_crps": 3.881798,
-      "mean_mae": 1.545328,
+      "mean_mae": 4.378282,
-      "bucket_hit_rate": 0.352941
+      "bucket_hit_rate": 0.225352
     },
     "delta": {
-      "crps": 0.161791,
+      "crps": 0.272913,
-      "mae": 0.301798,
+      "mae": 0.563211,
-      "bucket_hit_rate": -0.235294
+      "bucket_hit_rate": -0.267606
     }
   },
   "by_city": {
     "ankara": {
-      "samples": 3,
+      "samples": 6,
-      "legacy_mean_crps": 2.707176,
+      "legacy_mean_crps": 1.365135,
-      "emos_mean_crps": 2.632737,
+      "emos_mean_crps": 1.454394,
-      "legacy_mean_mae": 3.0,
+      "legacy_mean_mae": 1.466667,
-      "emos_mean_mae": 2.90439,
+      "emos_mean_mae": 1.781865,
-      "legacy_bucket_hit_rate": 0.333333,
+      "legacy_bucket_hit_rate": 0.666667,
       "emos_bucket_hit_rate": 0.333333
     },
+    "atlanta": {
+      "samples": 2,
+      "legacy_mean_crps": 30.449382,
+      "emos_mean_crps": 30.2136,
+      "legacy_mean_mae": 32.015,
+      "emos_mean_mae": 31.431544,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "buenos aires": {
+      "samples": 2,
+      "legacy_mean_crps": 9.113412,
+      "emos_mean_crps": 9.25091,
+      "legacy_mean_mae": 10.27,
+      "emos_mean_mae": 10.234011,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
     "busan": {
       "samples": 1,
-      "legacy_mean_crps": 0.264698,
+      "legacy_mean_crps": 0.451701,
-      "emos_mean_crps": 0.797946,
+      "emos_mean_crps": 1.176676,
       "legacy_mean_mae": 0.3,
-      "emos_mean_mae": 1.213185,
+      "emos_mean_mae": 1.887548,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 0.0
     },
     "chengdu": {
       "samples": 1,
-      "legacy_mean_crps": 0.218827,
+      "legacy_mean_crps": 0.306469,
-      "emos_mean_crps": 0.321502,
+      "emos_mean_crps": 0.555327,
       "legacy_mean_mae": 0.3,
-      "emos_mean_mae": 0.539971,
+      "emos_mean_mae": 0.929502,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "chicago": {
+      "samples": 1,
+      "legacy_mean_crps": 1.250268,
+      "emos_mean_crps": 2.090175,
+      "legacy_mean_mae": 0.0,
+      "emos_mean_mae": 3.396557,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "dallas": {
+      "samples": 1,
+      "legacy_mean_crps": 2.173363,
+      "emos_mean_crps": 1.05572,
+      "legacy_mean_mae": 0.0,
+      "emos_mean_mae": 1.653475,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 0.0
     },
     "hong kong": {
-      "samples": 3,
+      "samples": 7,
-      "legacy_mean_crps": 0.321946,
+      "legacy_mean_crps": 0.706099,
-      "emos_mean_crps": 0.419198,
+      "emos_mean_crps": 1.573367,
-      "legacy_mean_mae": 0.27,
+      "legacy_mean_mae": 0.657143,
-      "emos_mean_mae": 0.673605,
+      "emos_mean_mae": 1.88803,
-      "legacy_bucket_hit_rate": 1.0,
+      "legacy_bucket_hit_rate": 0.714286,
-      "emos_bucket_hit_rate": 0.666667
+      "emos_bucket_hit_rate": 0.142857
     },
     "istanbul": {
       "samples": 3,
-      "legacy_mean_crps": 1.308328,
+      "legacy_mean_crps": 1.264103,
-      "emos_mean_crps": 1.231189,
+      "emos_mean_crps": 1.266844,
-      "legacy_mean_mae": 1.38,
+      "legacy_mean_mae": 1.5,
-      "emos_mean_mae": 1.211693,
+      "emos_mean_mae": 1.415758,
-      "legacy_bucket_hit_rate": 0.666667,
+      "legacy_bucket_hit_rate": 0.333333,
       "emos_bucket_hit_rate": 0.666667
     },
-    "seoul": {
+    "london": {
-      "samples": 1,
-      "legacy_mean_crps": 0.208763,
-      "emos_mean_crps": 2.708973,
-      "legacy_mean_mae": 0.3,
-      "emos_mean_mae": 3.02868,
-      "legacy_bucket_hit_rate": 1.0,
-      "emos_bucket_hit_rate": 0.0
-    },
-    "taipei": {
       "samples": 2,
-      "legacy_mean_crps": 0.767971,
+      "legacy_mean_crps": 3.885033,
-      "emos_mean_crps": 1.205469,
+      "emos_mean_crps": 4.057521,
-      "legacy_mean_mae": 1.195,
+      "legacy_mean_mae": 4.135,
-      "emos_mean_mae": 1.937374,
+      "emos_mean_mae": 4.313103,
       "legacy_bucket_hit_rate": 0.0,
       "emos_bucket_hit_rate": 0.0
     },
-    "tokyo": {
+    "lucknow": {
+      "samples": 2,
+      "legacy_mean_crps": 2.487193,
+      "emos_mean_crps": 2.021264,
+      "legacy_mean_mae": 3.205,
+      "emos_mean_mae": 2.47462,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.5
+    },
+    "madrid": {
+      "samples": 2,
+      "legacy_mean_crps": 6.27726,
+      "emos_mean_crps": 6.588692,
+      "legacy_mean_mae": 7.33,
+      "emos_mean_mae": 7.486713,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "miami": {
+      "samples": 2,
+      "legacy_mean_crps": 28.637631,
+      "emos_mean_crps": 28.800796,
+      "legacy_mean_mae": 30.175,
+      "emos_mean_mae": 30.107593,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "milan": {
       "samples": 3,
-      "legacy_mean_crps": 1.129263,
+      "legacy_mean_crps": 4.401392,
-      "emos_mean_crps": 0.763363,
+      "emos_mean_crps": 4.947416,
-      "legacy_mean_mae": 1.3,
+      "legacy_mean_mae": 4.06,
-      "emos_mean_mae": 1.081641,
+      "emos_mean_mae": 5.6958,
-      "legacy_bucket_hit_rate": 0.333333,
+      "legacy_bucket_hit_rate": 0.666667,
       "emos_bucket_hit_rate": 0.333333
+    },
+    "munich": {
+      "samples": 2,
+      "legacy_mean_crps": 3.145192,
+      "emos_mean_crps": 3.200946,
+      "legacy_mean_mae": 3.64,
+      "emos_mean_mae": 3.634475,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "new york": {
+      "samples": 1,
+      "legacy_mean_crps": 3.692845,
+      "emos_mean_crps": 3.881457,
+      "legacy_mean_mae": 4.94,
+      "emos_mean_mae": 4.954702,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "paris": {
+      "samples": 2,
+      "legacy_mean_crps": 4.013782,
+      "emos_mean_crps": 4.130614,
+      "legacy_mean_mae": 4.265,
+      "emos_mean_mae": 4.314216,
+      "legacy_bucket_hit_rate": 0.5,
+      "emos_bucket_hit_rate": 0.5
+    },
+    "sao paulo": {
+      "samples": 2,
+      "legacy_mean_crps": 5.540967,
+      "emos_mean_crps": 5.483258,
+      "legacy_mean_mae": 6.57,
+      "emos_mean_mae": 6.358434,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "seattle": {
+      "samples": 1,
+      "legacy_mean_crps": 0.315488,
+      "emos_mean_crps": 0.271998,
+      "legacy_mean_mae": 0.0,
+      "emos_mean_mae": 0.105037,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 1.0
+    },
+    "seoul": {
+      "samples": 3,
+      "legacy_mean_crps": 0.508331,
+      "emos_mean_crps": 1.718042,
+      "legacy_mean_mae": 0.2,
+      "emos_mean_mae": 2.57121,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "shanghai": {
+      "samples": 3,
+      "legacy_mean_crps": 0.299116,
+      "emos_mean_crps": 0.553444,
+      "legacy_mean_mae": 0.1,
+      "emos_mean_mae": 0.71746,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 0.666667
+    },
+    "shenzhen": {
+      "samples": 1,
+      "legacy_mean_crps": 1.198351,
+      "emos_mean_crps": 0.507942,
+      "legacy_mean_mae": 1.3,
+      "emos_mean_mae": 0.644561,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "singapore": {
+      "samples": 2,
+      "legacy_mean_crps": 0.281993,
+      "emos_mean_crps": 0.877023,
+      "legacy_mean_mae": 0.15,
+      "emos_mean_mae": 1.142765,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 0.5
+    },
+    "taipei": {
+      "samples": 5,
+      "legacy_mean_crps": 0.950739,
+      "emos_mean_crps": 1.257972,
+      "legacy_mean_mae": 0.94,
+      "emos_mean_mae": 1.687014,
+      "legacy_bucket_hit_rate": 0.6,
+      "emos_bucket_hit_rate": 0.2
+    },
+    "tel aviv": {
+      "samples": 2,
+      "legacy_mean_crps": 0.446758,
+      "emos_mean_crps": 0.482341,
+      "legacy_mean_mae": 0.3,
+      "emos_mean_mae": 0.62659,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 0.5
+    },
+    "tokyo": {
+      "samples": 5,
+      "legacy_mean_crps": 0.879366,
+      "emos_mean_crps": 0.989336,
+      "legacy_mean_mae": 1.022,
+      "emos_mean_mae": 1.450474,
+      "legacy_bucket_hit_rate": 0.2,
+      "emos_bucket_hit_rate": 0.2
+    },
+    "toronto": {
+      "samples": 2,
+      "legacy_mean_crps": 5.497916,
+      "emos_mean_crps": 6.296632,
+      "legacy_mean_mae": 6.33,
+      "emos_mean_mae": 7.003984,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "warsaw": {
+      "samples": 3,
+      "legacy_mean_crps": 1.618875,
+      "emos_mean_crps": 1.649515,
+      "legacy_mean_mae": 2.056667,
+      "emos_mean_mae": 2.073494,
+      "legacy_bucket_hit_rate": 0.333333,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "wellington": {
+      "samples": 2,
+      "legacy_mean_crps": 0.364919,
+      "emos_mean_crps": 0.939463,
+      "legacy_mean_mae": 0.15,
+      "emos_mean_mae": 1.007263,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 0.5
     }
   }
 }

scripts/evaluate_probability_calibration.py

@@ -18,9 +18,14 @@ from src.analysis.settlement_rounding import apply_city_settlement  # noqa: E402
 from scripts.fit_probability_calibration import (  # noqa: E402
     _default_history_arg,
     _extract_samples,
+    _load_history_with_fallback,
     _load_json_if_exists,
+    _load_legacy_training_samples,
+    _load_snapshot_rows,
+    _load_training_feature_history,
+    _load_truth_history,
+    merge_samples_with_legacy_archive,
 )
-from src.analysis.deb_algorithm import load_history  # noqa: E402
 
 
 def _mean(values):
@@ -93,12 +98,20 @@ def main():
     )
     args = parser.parse_args()
 
-    history = load_history(args.history_file)
+    history = _load_history_with_fallback(args.history_file)
+    training_feature_history = _load_training_feature_history()
+    truth_history = _load_truth_history()
     settlement_history = _load_json_if_exists(args.settlement_history)
+    snapshot_rows = _load_snapshot_rows(None)
+    legacy_training_samples = _load_legacy_training_samples()
     samples, filled_actual_from_history = _extract_samples(
         history,
+        training_feature_history=training_feature_history,
+        truth_history=truth_history,
         settlement_history=settlement_history,
+        snapshot_rows=snapshot_rows,
     )
+    samples = merge_samples_with_legacy_archive(samples, legacy_training_samples)
 
     legacy_crps = []
     emos_crps = []

scripts/fit_probability_calibration.py

@@ -42,6 +42,15 @@ def _load_json_if_exists(path):
     return data if isinstance(data, dict) else {}
 
 
+def _legacy_training_samples_path():
+    return os.path.join(
+        PROJECT_ROOT,
+        "artifacts",
+        "probability_calibration",
+        "training_samples.json",
+    )
+
+
 def _legacy_history_path():
     return os.path.join(PROJECT_ROOT, "data", "daily_records.json")
 
@@ -107,6 +116,14 @@ def _load_snapshot_rows(path):
     return rows
 
 
+def _load_legacy_training_samples(path=None):
+    payload = _load_json_if_exists(path or _legacy_training_samples_path())
+    rows = payload.get("samples") if isinstance(payload, dict) else None
+    if not isinstance(rows, list):
+        return []
+    return [row for row in rows if isinstance(row, dict)]
+
+
 def _actual_high_for(history, truth_history, settlement_history, city, date_str):
     city_rows = (history or {}).get(city) or {}
     record = city_rows.get(date_str) or {}
@@ -327,6 +344,40 @@ def _extract_samples(history, training_feature_history=None, truth_history=None,
     return snapshot_samples + daily_samples, snapshot_filled + daily_filled
 
 
+def merge_samples_with_legacy_archive(samples, legacy_samples=None):
+    merged = []
+    seen = set()
+    for sample in samples or []:
+        if not isinstance(sample, dict):
+            continue
+        key = (
+            str(sample.get("city") or "").strip().lower(),
+            str(sample.get("date") or "").strip(),
+            str(sample.get("sample_source") or "").strip().lower(),
+        )
+        if not key[0] or not key[1]:
+            continue
+        if key in seen:
+            continue
+        merged.append(sample)
+        seen.add(key)
+    for sample in legacy_samples or []:
+        if not isinstance(sample, dict):
+            continue
+        key = (
+            str(sample.get("city") or "").strip().lower(),
+            str(sample.get("date") or "").strip(),
+            str(sample.get("sample_source") or "").strip().lower(),
+        )
+        if not key[0] or not key[1]:
+            continue
+        if key in seen:
+            continue
+        merged.append(sample)
+        seen.add(key)
+    return merged
+
+
 def main():
     parser = argparse.ArgumentParser(description="Fit PolyWeather probability calibration parameters.")
     parser.add_argument(
@@ -366,6 +417,7 @@ def main():
     truth_history = _load_truth_history()
     settlement_history = _load_json_if_exists(args.settlement_history)
     snapshot_rows = _load_snapshot_rows(args.snapshot_file)
+    legacy_training_samples = _load_legacy_training_samples()
     samples, filled_actual_from_history = _extract_samples(
         history,
         training_feature_history=training_feature_history,
@@ -373,6 +425,7 @@ def main():
         settlement_history=settlement_history,
         snapshot_rows=snapshot_rows,
     )
+    samples = merge_samples_with_legacy_archive(samples, legacy_training_samples)
     calibration = fit_calibration(samples, version=args.version)
     if not samples:
         calibration = default_calibration_payload(
@@ -382,6 +435,7 @@ def main():
     calibration.setdefault("metrics", {})
     calibration["metrics"]["filled_actual_from_history"] = filled_actual_from_history
     calibration["metrics"]["settlement_history_city_count"] = len(settlement_history)
+    calibration["metrics"]["legacy_archive_samples"] = len(legacy_training_samples)
     try:
         calibration["source"] = os.path.relpath(args.output, PROJECT_ROOT)
     except ValueError:

src/analysis/deb_algorithm.py

@@ -882,8 +882,8 @@ def update_daily_record(
     except Exception as e:
         logger.error(f"Error persisting training feature record city={city_name} date={date_str}: {e}")
 
-    # 自动清理：只保留最近 14 天的记录（DEB 只用 7 天，14 天留足余量）
+    # 自动清理：训练特征需要更长窗口，保留最近 180 天的记录
-    cutoff = (datetime.now() - timedelta(days=14)).strftime("%Y-%m-%d")
+    cutoff = (datetime.now() - timedelta(days=180)).strftime("%Y-%m-%d")
     for city in list(data.keys()):
         old_dates = [d for d in data[city] if d < cutoff]
         for d in old_dates:
@@ -892,7 +892,7 @@ def update_daily_record(
     if mode == STATE_STORAGE_SQLITE:
         try:
             _daily_record_repo.upsert_record(city_name, date_str, existing)
-            cutoff = (datetime.now() - timedelta(days=14)).strftime("%Y-%m-%d")
+            cutoff = (datetime.now() - timedelta(days=180)).strftime("%Y-%m-%d")
             _daily_record_repo.delete_older_than(cutoff)
         except Exception as e:
             logger.error(f"Error upserting daily record to sqlite city={city_name} date={date_str}: {e}")